BGFANet: boundary-guided fusion attention network for polyp segmentation in colonoscopy images

Abstract

Precise segmentation of colorectal polyps in colonoscopy images plays a vital role in early cancer screening and clinical decision-making. However, the task remains challenging due to large variations in polyp size and shape, ambiguous and low-contrast boundaries, and noise interference in colonoscopic images. These factors often lead to inaccurate localization and incomplete segmentation, especially for small polyps that exhibit subtle visual distinctions from surrounding tissues. To overcome these challenges, we propose the boundary-guided fusion attention network (BGFANet) for accurate polyp segmentation. Specifically, we first introduce a deep fusion guidance module that effectively aggregates high-level semantic features to generate initial foreground and background guidance maps for preliminary polyp localization. Secondly, the multi-scale edge attention aggregation module integrates multi-scale Gaussian filtering with gradient extraction to suppress background noise and enhance boundary representations, thereby improving the segmentation of polyps with ambiguous boundaries. Finally, a novel boundary-guided fusion attention (BGFA) module is proposed to effectively integrate cross-level features and exploit both local and global contextual information. Extensive experiments conducted on five challenging public datasets demonstrate the superiority of BGFANet over existing mainstream methods. Specifically, BGFANet achieves a leading mDice of 0.906 on CVC–300, and outperforms other state-of-the-art methods on CVC-ColonDB and ETIS-LaribPolypDB by absolute gains of 6.2% and 4.3% in mDice, respectively.